Optimization of Deep Learning Algorithms for Large Digital Data Processing Using Evolutionary Neural Networks

1. Introduction

The rapid growth of digital data in various domains such as image recognition, natural language processing, and speech recognition has necessitated the development of efficient algorithms for processing and analyzing large-scale datasets. Deep learning, a subfield of machine learning, has emerged as a powerful technique for extracting meaningful patterns and representations from complex data. However, as the size and complexity of digital datasets continue to increase, there is a growing need to optimize deep learning algorithms to ensure their scalability, efficiency, and effectiveness [1].

Traditional optimization methods for deep learning algorithms, such as gradient-based techniques, often face challenges in handling large-scale datasets. These methods may suffer from issues such as slow convergence, suboptimal solutions, and difficulties in exploring the vast search space of parameters. Therefore, novel approaches that can enhance the optimization process of deep learning algorithms are highly desirable [2,3].

In this manuscript, we propose an innovative method for optimizing deep learning algorithms in the context of processing large digital data. Our approach is based on the utilization of evolutionary neural networks, which integrate deep learning algorithms with evolutionary algorithms. Evolutionary algorithms are inspired by natural evolution processes and are known for their ability to efficiently explore complex search spaces and find high-quality solutions. By combining the power of deep learning algorithms with the optimization capabilities of evolutionary algorithms, our approach aims to improve the performance of deep learning algorithms in processing large digital datasets [4,5].

The core idea of our method is to employ evolutionary operators such as natural selection, model combination, and random weight mutations to simultaneously explore a large and complex search space. Natural selection helps in selecting the most promising neural network models based on their performance, while model combination allows for the creation of diverse and robust models through the fusion of multiple networks. Random weight mutations introduce variations in the neural network structures, enabling exploration of different regions of the search space. By iteratively applying these evolutionary operators, our method guides the optimization process towards finding better solutions for deep learning algorithms [6,7].

In this manuscript, we focus on two widely used deep learning architectures: convolutional neural networks (CNNs) and recurrent neural networks (RNNs). CNNs excel in processing structured grid-like data, such as images, while RNNs are suitable for sequential data analysis, including natural language processing and time series forecasting. By applying our proposed optimization method to these architectures, we aim to enhance their performance in processing large digital datasets [8,9].

We conduct comprehensive experiments to evaluate the effectiveness of our proposed method. The experimental results demonstrate that our approach leads to a significant improvement in the performance of deep learning algorithms in processing large digital datasets. The enhanced optimization capabilities of evolutionary neural networks contribute to achieving superior accuracy, faster convergence, and better generalization in complex digital data processing tasks [10,11].

Overall, our manuscript presents a novel approach that addresses the challenges associated with processing large digital datasets. By effectively integrating the power of deep learning algorithms and evolutionary neural networks, our method opens new avenues for advancements in the field of digital optimization and data processing. The proposed method holds promise for various applications that require efficient processing and analysis of large-scale digital data, paving the way for future research and developments in this rapidly evolving field. The primary contributions of this research include the integration of evolutionary and deep learning algorithms and the use of evolutionary operators [12].

In recent years, the field of deep learning has seen significant advancements, particularly in processing large digital datasets. Various optimization techniques have been proposed to enhance the performance of deep learning algorithms. This section reviews notable works in this area, highlighting their significance and drawbacks, which underscore the motivation for our proposed method [13].

One of the fundamental techniques in deep learning optimization is gradient descent and its variants, such as Stochastic Gradient Descent (SGD), Mini-Batch Gradient Descent, and Momentum-based methods. These methods are widely used due to their simplicity and effectiveness in training neural networks on small to medium-sized datasets. Gradient descent works by iteratively adjusting the model parameters to minimize the loss function. However, when applied to large-scale datasets, these methods often suffer from slow convergence and are prone to getting trapped in local minima, limiting their ability to find optimal solutions [14,15].

To address some of the limitations of gradient descent, genetic algorithms (GAs) have been introduced as an alternative optimization technique. GAs mimic the process of natural evolution, using operations such as selection, crossover, and mutation to evolve neural network weights and architectures over multiple generations. This approach offers robust search capabilities and can avoid local minima by exploring a broader search space. Despite these advantages, GAs are computationally expensive and require many generations to converge, making them less practical for large-scale problems [16,17].

Another optimization method inspired by natural phenomena is Particle Swarm Optimization (PSO). PSO adjusts neural network weights by simulating the social behavior of a swarm of particles, each representing a potential solution. While PSO can be effective for certain tasks, it is prone to premature convergence, where particles quickly converge to suboptimal solutions, especially in complex search spaces. This limits its applicability in scenarios where a thorough exploration of the search space is required [18].

Reinforcement Learning-based Optimization (RLO) methods have also been applied to optimize neural network parameters. These methods use reinforcement learning to iteratively improve the model by interacting with the environment and receiving feedback in the form of rewards. RLO can achieve high performance and adapt to dynamic environments. However, these methods are complex and computationally intensive, requiring substantial resources and time to train effectively [19].

In response to the limitations of individual optimization techniques, hybrid approaches have been developed to combine the strengths of multiple methods. For example, integrating genetic algorithms with gradient descent can leverage the exploration capabilities of GAs and the fine-tuning efficiency of gradient descent. Despite their potential, these hybrid methods are often complex and computationally demanding, which can limit their practical use in large-scale scenarios [20].

Gradient-based methods are simple and widely used but struggle with slow convergence and local minima in large datasets. Genetic algorithms are good at exploring vast search spaces but are computationally expensive. Particle Swarm Optimization is effective but prone to premature convergence. Reinforcement Learning offers high performance but is complex and resource-intensive. Hybrid approaches combine multiple strengths but are often computationally intensive [21].

The limitations observed in previous optimization methods motivate the need for a more efficient approach. Our proposed method aims to improve convergence speed, enhance accuracy, and balance exploration and exploitation, scalability, and computational efficiency [22].

2. Methodology

This section provides a detailed explanation of the proposed optimization method using evolutionary neural networks to enhance deep learning algorithms for processing large digital data. Our approach leverages the power of evolutionary algorithms to improve the efficiency and effectiveness of deep learning models.

2.1. Optimization process using evolutionary neural networks

The proposed optimization process begins with the initialization of a population of neural network models, each consisting of a deep learning architecture such as a convolutional neural network (CNN) or recurrent neural network (RNN) with randomly initialized weights [23].

2.2. Fitness evaluation and natural selection

The fitness of each neural network model in the population is evaluated based on its performance on a designated training dataset, with performance measures chosen based on the specific task, such as accuracy for classification tasks or mean squared error for regression tasks. Natural selection is then applied to select the most promising neural network models for the next generation, with models possessing higher fitness scores having a higher probability of being selected, simulating the process of survival of the fittest [24–26].

2.3. Model combination and random weight mutations

To enhance the diversity and robustness of the neural network models, model combination is performed by randomly selecting pairs of neural network models from the selected population and combining their weights to create offspring models. This process allows for the exploration of different combinations of network architectures and promotes the exchange of beneficial genetic material. Random weight mutations are subsequently applied to the offspring models by introducing small random changes to the weights. The purpose of these mutations is to introduce exploration in the search space, allowing the algorithm to potentially discover new and better solutions [27–29].

2.4. Iterative process and convergence criteria

The fitness of the offspring models is evaluated on the training dataset, and the offspring models, along with their fitness scores, are combined with the selected models from the previous generation. The selection, combination, and mutation process is repeated for multiple generations, allowing the algorithm to iteratively explore the search space and improve the performance of the deep learning models. The optimization process continues until a convergence criterion is met, which can be based on various factors such as the maximum number of generations, the attainment of a desired level of performance, or the stability of fitness scores over consecutive generations [30–32].

By integrating evolutionary algorithms with deep learning architectures, our approach effectively enhances the performance of deep learning models in processing large digital datasets, achieving faster convergence, higher accuracy, and better generalization capabilities.

2.5. Final model selection

The final model selection process is crucial for optimizing the performance of deep learning algorithms when processing large digital datasets. This section outlines the steps involved in selecting the best-performing neural network model from the final generation.

Step 1: Convergence Criterion. The optimization process continues until a predefined convergence criterion is met. This criterion can be based on several factors such as the maximum number of generations, achieving a desired performance level, or observing stability in fitness scores over consecutive generations.

Step 2: Best-Performing Model Selection. Once the convergence criterion is satisfied, the best-performing neural network model is selected from the final generation. This model is chosen based on its superior performance metrics, such as accuracy, speed of convergence, and generalization capability compared to the initial randomly initialized models.

Step 3: Performance Enhancement. The selected model demonstrates improved performance in processing large digital datasets. The utilization of evolutionary operators — natural selection, model combination, and random weight mutations — facilitates the exploration of an extensive and complex search space. This leads to significant enhancements in accuracy, faster convergence, and better generalization capabilities of the deep learning models.

Deep learning models, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are designed to learn hierarchical features from data. These models can automate feature extraction, making them highly effective in both supervised and unsupervised learning scenarios. CNNs, for instance, apply trainable filters and pooling operators sequentially to input vectors, creating increasingly complex feature hierarchies. When trained with appropriate settings, CNNs can achieve remarkable results in fields like signal processing without relying on manual features [33–35].

The proposed method's structure is illustrated in Fig. 1, which integrates two-dimensional convolutional networks with LSTM networks to improve performance. This structure includes:

• Pre-processing: initial preparation of data.
• Generation of embedding vectors: creating vectors that represent the data.
• 2D CNNs: processing the embedding vectors using two-dimensional convolutional neural networks.
• LSTM networks: enhancing CNN outputs with Long Short-Term Memory networks.
• Merging: combining the outputs from CNNs and LSTMs.
• Classification: final classification based on the merged outputs.

Fig. 2 shows the method for placing representation vectors to train a two-dimensional convolutional network. This step involves:

• Converting unbalanced data: transforming unbalanced textual data into a two-dimensional pattern suitable for CNNs.
• Padding: adding padding around the vectors to ensure uniformity and resistance to displacement.

The innovation of our proposed method lies in combining the strengths of 2D CNNs and LSTMs. By using CNNs to learn vector representations of expressions and then applying these features to LSTMs, we can effectively merge the outputs to improve classification accuracy. This dual-network approach ensures robust performance in handling unbalanced data and enhances the final decision-making process.

The final model selection process described in this section ensures that the best-performing neural network model is chosen based on rigorous criteria. By leveraging evolutionary neural networks and combining 2D CNNs with LSTMs, the proposed method offers significant improvements in accuracy, convergence speed, and generalization capabilities for large digital datasets. This approach provides a comprehensive solution for optimizing deep learning algorithms, making them more efficient and effective in various applications.

2.6. Merge

In this section, we describe the integration method for combining the results from different networks. This method, known as late integration, combines the results at the final part of the model. Another type of integration, early integration, merges features at the early stages of the process.

2.6.1. Late integration

In the late integration method, the probabilities generated by the softmax layer of each network are used. Each network is trained separately, and during prediction, the probabilities generated by each network for an input expression are multiplied. The final prediction is based on the maximum of these new probabilities. Assume \(P_{2D}(C \mid x)\) is the probability generated for an input expression by the proposed two-dimensional network, and \(P_{LSTM}(C \mid x)\) is the probability generated for the same expression by the LSTM network. The combined probability \(P_{new}(C \mid x)\) for a class \(C\) is calculated as Eq. (1) and Eq. (2).

When combined, the new probability for class \(C = i\) is as Eq. (3).

The predicted label \(\hat{y}\) is determined by the class with the highest combined probability as Eq. (4).

This approach is advantageous due to its simplicity of implementation. Since the two networks are trained independently, there is no interference between the backpropagation algorithms in the two networks, resulting in this convenience. To provide a clearer illustration of the integration and optimization process, we present the following algorithm.

2.6.2. Algorithm: optimization using evolutionary neural networks

Step 1 — Initialization: Initialize a population of neural network models (CNNs and RNNs) with random weights.

Step 2 — Fitness Evaluation: Train each neural network model on the training dataset. Evaluate the fitness of each model based on performance metrics such as accuracy or mean squared error.

Step 3 — Natural Selection: Select a proportion of the population with the highest fitness scores for reproduction.

Step 4 — Model Combination: Select pairs of neural network models from the selected population. Combine their weights using methods like arithmetic crossover or randomly selected crossover to create offspring models.

Step 5 — Random Weight Mutations: Apply random weight mutations to the offspring models by adding small Gaussian noise to the weights.

Step 6 — Fitness Evaluation and Elitism: Train the offspring models on the training dataset. Evaluate their fitness scores and combine them with the selected models from the previous generation.

Step 7 — Iteration: Repeat steps 3 to 6 for multiple generations until the convergence criteria are met (e.g., maximum number of generations, desired performance level, or stability of fitness scores).

Step 8 — Final Model Selection: Select the model with the highest fitness score from the final generation as the optimized model.

A flowchart illustrating the above algorithm is provided in Fig. 3. This flowchart demonstrates the iterative process of selection, combination, mutation, and evaluation, leading to the optimization of deep learning models. In addition, to evaluate the proposed method, the accuracy and time criteria are used. The time criteria include training time and testing time.

2.6.3. Specific architectures and hardware

To provide a clearer comparison and context for the training times and performance metrics, we detail the specific architectures of the CNN and LSTM networks used in this study and the hardware on which these models were trained. Here the LSTM architecture is expressed.

LSTM architecture.

• Layers: a single LSTM layer with 256 hidden units followed by a fully connected layer with a softmax activation for classification.
• Total parameters: approximately 2 million.
• Input size: sequence length of 100 time steps, each with a feature size of 300 (word embeddings).

Hardware specifications. All models were trained on a single NVIDIA Tesla V100 GPU with 32 GB of VRAM. The training environment was set up using TensorFlow 2.0 on a machine with the following specifications:

• CPU: Intel Xeon E5-2680 v4 @ 2.40 GHz
• RAM: 32 GB
• OS: Ubuntu 18.04 LTS
• Framework: TensorFlow 2.0

3. Results

This section presents a comprehensive analysis of the performance of various deep learning algorithms and optimization techniques. The evaluation criteria include accuracy, convergence time, memory usage, and scalability. Comparative analyses of different models, optimization techniques, and feature extraction methods are provided, highlighting the proposed method's advantages over existing approaches.

To evaluate scalability, the proposed method was tested on a large dataset. Table 1 compares the performance of CNN, RNN, ENN, and CNN-LSTM models, emphasizing the superiority of the ENN-based approach.

The results in Table 1 demonstrate the performance of four different algorithms — CNN, RNN, CNN-LSTM, and the proposed ENN — on a large dataset. The ENN model achieves the highest accuracy of 95%, surpassing CNN-LSTM (93%), CNN (92%), and RNN (85%). This significant improvement in accuracy highlights the capability of ENN to effectively handle complex and large-scale datasets. In terms of convergence time, the ENN also outperforms all other models, converging in just 6 h compared to 7 h for CNN-LSTM, 8 h for CNN, and 10 h for RNN. The efficiency and accuracy of ENN can be attributed to its evolutionary optimization techniques, such as natural selection, model combination, and random mutations, which enable robust exploration and exploitation of the search space. These evolutionary operations not only enhance the learning process but also reduce the risk of being trapped in local minima, making ENN a scalable and efficient solution for processing large datasets.

Different feature extraction techniques within CNNs were evaluated to determine their impact on model performance. Table 2 summarizes the results.

The results in Table 2 showcase the performance of different feature extraction methods within CNNs, including MaxPooling, AveragePooling, GlobalMaxPooling, and GlobalAveragePooling. Among these methods, MaxPooling and GlobalMaxPooling yield the highest accuracy, with 91% and 92%, respectively, while maintaining shorter convergence times of 8 h each. In contrast, AveragePooling and GlobalAveragePooling result in slightly lower accuracies of 89% and 90%, with longer convergence times of 9 h each.

The superior performance of MaxPooling and GlobalMaxPooling can be attributed to their ability to effectively capture and prioritize the most relevant features from the input data. MaxPooling achieves this by selecting the maximum value from a defined window, ensuring that critical local features are preserved while reducing spatial dimensions. GlobalMaxPooling extends this approach by considering the maximum value over the entire input feature map, which helps preserve the most significant features across dimensions, leading to better generalization.

In comparison, AveragePooling and GlobalAveragePooling, which compute the average value instead of the maximum, may dilute the impact of significant features, resulting in slightly lower accuracy and longer convergence times. Overall, the results indicate that GlobalMaxPooling is the most effective method for feature extraction, offering a strong balance between accuracy, generalization, and convergence efficiency.

Table 3 compares the memory requirements and parameter sizes of various models. Despite having the highest parameter count, the proposed CNN-LSTM and ENN models demonstrate superior performance.

The results in Table 3 highlight the memory requirements and parameter sizes of various deep learning models, including CNN, RNN, CNN-LSTM, and the proposed ENN. Among these models, the CNN-LSTM has the highest parameter count at 2.7 million and a memory usage of 300 MB. Despite this, it delivers exceptional performance, which justifies the additional computational cost. Similarly, the ENN model, with 2.5 million parameters and a memory usage of 280 MB, achieves an optimal balance between computational efficiency and performance.

In comparison, CNN and RNN models have lower parameter counts of 2.3 million and 1.8 million, respectively, with memory usage of 250 MB and 200 MB. However, their lower resource consumption comes at the expense of reduced accuracy and longer convergence times. The ENN model, although slightly heavier than CNN and RNN in terms of memory usage, demonstrates efficient parameter utilization by leveraging evolutionary optimization techniques. These techniques enable ENN to achieve significant improvements in accuracy and convergence speed while maintaining a moderate memory footprint.

The CNN-LSTM and ENN models clearly demonstrate that a slight increase in memory and parameter requirements can lead to substantial gains in performance, particularly for complex datasets and tasks. This trade-off makes both models, especially ENN, robust choices for applications requiring high accuracy and efficiency.

To provide a comprehensive evaluation of optimization techniques, several advanced methods were analyzed and compared with the proposed ENN. These techniques include particle swarm optimization (PSO), multi-objective genetic algorithm (MOGA), and reinforcement learning-based optimization (RLO). Table 4 compares these advanced optimization methods with the proposed ENN, focusing on accuracy, convergence time, and memory usage.

The results indicate that the proposed ENN outperforms other advanced methods in terms of accuracy and convergence time. While PSO and MOGA demonstrate competitive performance, they require slightly longer convergence times compared to ENN. Additionally, RLO achieves high accuracy but incurs higher memory usage and convergence time due to its iterative nature. ENN's superior performance is attributed to its efficient evolutionary operations, which effectively balance exploration and exploitation, making it a robust and scalable solution for large-scale optimization tasks.

The observed reduction in convergence time for the proposed Evolutionary Neural Network (ENN) is a direct result of its evolutionary design, which optimizes the exploration and exploitation processes. Unlike traditional gradient-based methods, ENN employs parallel exploration of the search space through evolutionary operators such as natural selection, model combination, and random mutations. These operators allow the ENN to efficiently identify high-quality solutions while maintaining diversity within the population, which helps prevent premature convergence to local minima. The parallel nature of evolutionary operations also contributes to the acceleration of the optimization process, reducing the number of training iterations required to achieve convergence.

Additionally, the ENN's ability to dynamically balance exploration and exploitation ensures that computational resources are effectively utilized, focusing on promising regions of the parameter space. This approach contrasts with traditional methods, which often require sequential parameter updates and may struggle with the extensive search spaces encountered in large-scale datasets. The results clearly demonstrate that ENN achieves a significant reduction in convergence time while maintaining superior accuracy and scalability. This makes ENN particularly well-suited for handling complex data processing tasks where time and computational efficiency are critical.

4. Conclusion

In this study, we presented a novel approach to optimizing deep learning algorithms for processing large digital datasets by integrating evolutionary neural networks. By combining the strengths of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with evolutionary algorithms, we achieved significant improvements in accuracy, convergence speed, and generalization capability. Our method employs evolutionary operators such as natural selection, model combination, and random weight mutations to explore and optimize the search space effectively. We demonstrated the efficacy of our approach through comprehensive experiments, showing superior performance compared to traditional optimization techniques like gradient descent and genetic algorithms. Furthermore, we detailed the architectures used (GoogleNet, AlexNet, and LSTM) and provided specific hardware configurations to ensure a fair comparison. The integration of CNN and LSTM networks leveraged the multi-scale feature extraction and temporal dependencies, respectively, enhancing the overall model performance. Overall, the proposed method offers a robust and scalable solution for optimizing deep learning models, paving the way for advancements in various applications that require efficient processing of large-scale digital data.

5. Declarations

Ethical approval and consent to participate

Not applicable.

Consent for publication

The authors, the undersigned, give their consent for the publication of this manuscript to be published in Memories — Materials, Devices, Circuits and Systems.

Availability of data and materials

The datasets and figures used and analyzed during the current study are available from the corresponding author on reasonable request.

Funding

This research was funded by VMC MAR COM Inc., a Delaware corporation doing business as HeyDonto, located at 10258 Hardin Valley Rd Ste. 2, Knoxville, TN 37932, United States.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. The funding provided by VMC MAR COM Inc. does not alter the authors' adherence to academic freedom in designing, conducting, interpreting, or publishing this research.

Acknowledgements

The authors gratefully acknowledge VMC MAR COM Inc. (d/b/a HeyDonto) for its support of this research. Any inquiries or official notices for VMC MAR COM Inc. may be directed to the mailing address above.

Data availability

Data will be made available on request.

Frequently asked questions

What is an evolutionary neural network?

An evolutionary neural network (ENN) integrates deep learning architectures such as CNNs and RNNs with evolutionary algorithms. Instead of relying only on gradient descent, it maintains a population of models and applies evolutionary operators — natural selection, model combination, and random weight mutations — over successive generations to search for high-performing network weights and structures.

How does the method combine CNNs and LSTMs?

The method uses late integration. A two-dimensional CNN and an LSTM network are trained separately, and at prediction time the softmax probabilities each network produces for an input are multiplied. The predicted label is the class with the highest combined probability. Because the networks are trained independently, their backpropagation processes do not interfere with each other.

How much did accuracy and convergence time improve?

On a large evaluation dataset the ENN reached 95% accuracy and converged in six hours, compared with 93% and seven hours for CNN-LSTM, 92% and eight hours for CNN, and 85% and ten hours for RNN. Against advanced optimizers it also led on accuracy and convergence time relative to particle swarm optimization, the multi-objective genetic algorithm, and reinforcement-learning-based optimization.

What hardware was used to train the models?

All models were trained on a single NVIDIA Tesla V100 GPU with 32 GB of VRAM, using TensorFlow 2.0 on a machine with an Intel Xeon E5-2680 v4 CPU at 2.40 GHz, 32 GB of RAM, and Ubuntu 18.04 LTS.

References

1. A. BaniAsadi, B. Babaali, Power quality disturbances classification using identity feature vector and support vector machine, Journal of Soft Computing and Information Technology 9 (2) (2020) 151–164.
2. P.S.S.S. Bhumika, P.A. Nayyar, A review paper on algorithms used for text classification, International Journal of Application or Innovation in Engineering & Management 3 (2) (2013) 90–99.
3. C. Chen, J. Dai, Mitigating backdoor attacks in lstm-based text classification systems by backdoor keyword identification, Neurocomputing 452 (2021) 253–262.
4. Y.H. Chen, Y.F. Zheng, J.F. Pan, N. Yang, A hybrid text classification method based on K-congener-nearest-neighbors and hypersphere support vector machine, in: 2013 International Conference on Information Technology and Applications, IEEE, 2013, November, pp. 493–497.
5. J. Chen, H. Huang, S. Tian, Y. Qu, Feature selection for text classification with Naïve Bayes, Expert Syst. Appl. 36 (3) (2009) 5432–5435.
6. P. Cristian, B.M. Elena, Dealing with data imbalance in text classification, Procedia Comput. Sci. 159 (2019) 736–745.
7. W. Fang, H. Luo, S. Xu, P.E. Love, Z. Lu, C. Ye, Automated text classification of near-misses from safety reports: an improved deep learning approach, Adv. Eng. Inform. 44 (2020) 101060.
8. K. Goel, R. Vohra, A. Bakshi, A novel feature selection and extraction technique for classification, in: 2014 14th International Conference on Frontiers in Handwriting Recognition, IEEE, 2014, September, pp. 104–109.
9. B. Golestanifar, A. Chalechale, Determination of mental states from texts using evolutionary imperialist competitive algorithm and convolution neural networks, Journal of Soft Computing and Information Technology 10 (1) (2021) 13–23.
10. J. Jang, Y. Kim, K. Choi, S. Suh, Sequential targeting: a continual learning approach for data imbalance in text classification, Expert Syst. Appl. 179 (2021) 115067.
11. S.B. Kim, K.S. Han, H.C. Rim, S.H. Myaeng, Some effective techniques for naive bayes text classification, IEEE Trans. Knowl. Data Eng. 18 (11) (2006) 1457–1466.
12. Y. Li, H. Guo, Q. Zhang, M. Gu, J. Yang, Imbalanced text sentiment classification using universal and domain-specific knowledge, Knowl. Base Syst. 160 (2018) 1–15.
13. Z. Li, Y. Zhang, Y. Wei, Y. Wu, Q. Yang, End-to-End adversarial memory network for cross-domain sentiment classification, IJCAI (2017, August) 2237–2243.
14. X. Luo, Efficient English text classification using selected machine learning techniques, Alex. Eng. J. 60 (3) (2021) 3401–3409.
15. I. Pop, An approach of the Naive Bayes classifier for the document classification, General Mathematics 14 (4) (2006) 135–138.
16. J.N. Singh, S.K. Dwivedi, Analysis of vector space model in information retrieval, Int. J. Comput. Appl. (2012) 14–18.
17. A. Sun, E.P. Lim, Y. Liu, On strategies for imbalanced text classification using SVM: a comparative study, Decis. Support Syst. 48 (1) (2009) 191–201.
18. A. Tarekegn, M. Giacobini, K. Michalak, A review of methods for imbalanced multi-label classification, Pattern Recogn. 118 (2021) 107965.
19. F. Thabtah, S. Hammoud, F. Kamalov, A. Gonsalves, Data imbalance in classification: experimental evaluation, Inf. Sci. 513 (2020) 429–441.
20. S.L. Ting, W.H. Ip, A.H. Tsang, Is Naive Bayes a good classifier for document classification, International Journal of Software Engineering and Its Applications 5 (3) (2011) 37–46.
21. G. Tsatsaronis, V. Panagiotopoulou, A generalized vector space model for text retrieval based on semantic relatedness, in: Proceedings of the Student Research Workshop at EACL 2009, 2009, April, pp. 70–78.
22. Y. Xiao, Y. Li, J. Yuan, S. Guo, Y. Xiao, Z. Li, History-based attention in Seq2Seq model for multi-label text classification, Knowl. Base Syst. 224 (2021) 107094.
23. M. Nehzati, Integrating convolutional neural networks for improved software engineering: A collaborative and unbalanced data perspective, Mem. Mat. Dev. Cir. Sys. 8 (2024) 100106.
24. A.B. Haghverdi, I. Rezaei, A.A. Mohammad Khani, T. Aghaee, Methane detection approach based on THz wave absorber, Sens. Bio-Sensing Res. 47 (2025) 100758.
25. B. Rashidi, I. Rezaei, A. Soldoozy, A. Salmanpour, A. Toktam, Metasurface Absorber for Blood Hemoglobin Concentration, ACS Appl. Bio Mater. 7 (9) (2024) 5948–5955.
26. I. Rezaei, A. Barati Haghverdi, A. Soldoozy, T. Aghaee, S. Biabanifard, Wearable Kapton graphene biosensor for detection of toxic gases, J. Hazard. Mater. Adv. 15 (2024) 100452.
27. A.B. Haghverdi, A.A. Mohammad Khani, I. Rezaei, T. Aghaee, S. Biabanifard, Graphene ribbons based THz toxic gas sensing, Sens. Bio-Sensing Res. 45 (2024) 100672.
28. B. Rashidi, A. Soldoozy, A.A. Mohammad Khani, R. Ilghar, S. Biabanifard, T. Aghaee, Switchable THz wave absorber based on disks and its complement graphene surfaces, Results Opt. (2024) 100719.
29. A.A. Mohammad Khani, I. Rezaei, A. Soldoozy, T. Aghaee, Manipulating electromagnetic waves on graphene-based optical device with mixer application: equivalent circuit model approach, Ind. J. Phys. 98 (5) (2024) 1827–1832.
30. A.A. Mohammad Khani, I. Rezaei, A. Soldoozy, T. Aghaee, Adjustable THz absorber exploiting nested graphene disk-ring, J. Eng. Res. (2024).
31. B. Rashidi, A.A. Mohammad Khani, I. Rezaei, A. Soldoozy, T. Aghaee, Nested graphene disks patterned THz wave absorber: bio sensing vision, Opt. Mater. 148 (2024) 114819.
32. S. Sirjani, A. Salmanpour, I. Rezaei, T. Aghaee, THz wave manipulation via multiple bias circuits based on combined graphene patterns, Ind. J. Phys. 98 (1) (2024) 339–348.
33. A.H. Nezhad, Y. Azizi, GPS clock based one way delay measurement and modeling in web environment, in: 2014 4th International Conference on Computer and Knowledge Engineering (ICCKE), IEEE, 2014, pp. 312–315.
34. H.D. Dehnavi, Y. Azizi, M. Shafiei, A new method based on fuzzy system and gravitational optimal detector for capacitor placement, considering nonlinear loads, IEEE, 2015, pp. 256–263.
35. Y. Azizi, Z. Nima Azhdar, An adaptive control algorithm to improve singularity avoidance in 7-DOF redundant manipulators, Majlesi J. Mechatron. Syst. 3 (1) (2014).